Slack ↔ WxO MCP Gateway

Author: Markus van Kempen
Email: mvankempen@ca.ibm.com · markus.van.kempen@gmail.com
Web: https://markusvankempen.github.io/ · GitHub

npm: @markusvankempen/slack-wxo-mcp-gateway · MCP: io.github.markusvankempen/slack-wxo-mcp-gateway

This GitHub repo is documentation + registry metadata. It does not include the runnable application source.
Install / run via npm: npx -y @markusvankempen/slack-wxo-mcp-gateway · Site: https://markusvankempen.github.io/

Pitch: MCP gateway that lifts watsonx Orchestrate Slack limitations — every-message wake-up, multi-channel→multi-agent routing, clean in-thread replies, and a streamable-http toolkit for WxO + Cursor / VS Code / Bob / Antigravity — without replacing your agents.

tags: wxo-limitations · byo-slack · every-message · multi-channel · multi-agent · thread-followups · gateway-thread · no-done-noise · mcp-toolkit · streamable-http · poller · code-engine · ngrok · agentic-ai

One config site: map many Slack channels → many WxO agents.
Poller (and optional Slack Events) wake agents.
Same host exposes an MCP toolkit (/mcp) for WxO / Cursor / other clients.

Deep dive: Why this MCP — lifting WxO limits

Architecture at a glance

flowchart LR
  subgraph Slack
    C1["#support"]
    C2["#orders"]
    C3["#ops"]
  end

  subgraph Gateway["Slack ↔ WxO MCP Gateway"]
    Bind["config.yaml bindings"]
    Poll["Poller / Events"]
    MCP["/mcp streamable-http"]
    UI["Admin UI /"]
  end

  subgraph WxO["watsonx Orchestrate"]
    A1["Agent A"]
    A2["Agent B"]
    A3["Agent C"]
  end

  subgraph Clients["MCP clients"]
    IDE["Cursor / VS Code / Bob / …"]
    TK["WxO toolkits"]
  end

  C1 & C2 & C3 --> Poll
  Poll --> Bind
  Bind --> A1 & A2 & A3
  A1 & A2 & A3 -.->|gateway_thread reply| Poll
  IDE & TK --> MCP
  UI --> Bind

---

Why this approach (WxO limits → lift)

WxO / Slack limit	Tag	Gateway lift
byo_slack ≈ @mention / DM only	every-message	Poller / Events wake agents on every human message
Hard to run many channels → many agents	multi-channel multi-agent	One bindings table + admin UI
Thread follow-ups easy to drop	thread-followups	Reads thread replies + context
Noisy finals (done, etc.) in Slack	gateway-thread no-done-noise	Gateway posts answers; filters noise
Agents need remote tools with real DNS	mcp-toolkit streamable-http	Hosted /mcp for Orchestrate toolkits
Ops stuck cloning pollers	ops-self-serve	MCP tools + diagnostics + logs
Slack ops only inside Slack/WxO UI	ide-parity	Same tools in Cursor, VS Code, Bob, Antigravity, Claude

WxO stays the brain (LLMs, skills, flows). This gateway is the Slack + routing + MCP edge.
Bring-your-own agent frameworks: docs/frameworks/ (LangGraph, LlamaIndex, OpenAI Agents).

---

npm / MCP identity

npm	@markusvankempen/slack-wxo-mcp-gateway
MCP name	io.github.markusvankempen/slack-wxo-mcp-gateway
Topics	mcp · mcp-server · slack · watsonx · watsonx-orchestrate · ibm · wxo · byo-slack · multi-channel · code-engine · streamable-http · cursor · agentic-ai

Full keyword list lives in package.json for npm discoverability.

---

Publish & run modes (A–D)

One package / one image — pick a mode (see docs/PUBLISH-MODES.md):

flowchart TB
  PKG["npm @markusvankempen/slack-wxo-mcp-gateway<br/>+ optional container image"]
  PKG --> A["A Local HTTP<br/>:3100 UI + /mcp + poller"]
  PKG --> B["B Podman / Docker<br/>:8080"]
  PKG --> C["C Code Engine<br/>HTTPS always-on"]
  PKG --> D["D IDE stdio<br/>Cursor / VS Code / Bob"]
  A --> N["ngrok demo tunnel"]
  A & B & C --> R["Remote /mcp clients"]
  D --> L["Local MCP session"]

Mode	Command	Use
A Local HTTP	./scripts/run.sh --mode http	UI + /mcp + poller on laptop
B Podman/Docker	./scripts/run.sh --mode podman	Same app in a container
C Code Engine	./scripts/run.sh --mode ce	Always-on HTTPS
D IDE MCP	./scripts/run.sh --mode ide	Cursor / VS Code stdio snippets (+ --exec)
Ngrok demo	./scripts/run.sh --mode ngrok	A + tunnel + WxO toolkit

./scripts/run.sh --mode ide      # print Cursor + VS Code mcp.json
./scripts/run.sh --mode http     # local host :3100
./scripts/run.sh --mode podman   # container :8080
./scripts/run.sh --mode ce       # IBM Code Engine

Deep guides: docs/local-ngrok/ · docs/code-engine/ · docs/ide/
Index: docs/README.md · Setup: SETUP.md

Copy-paste IDE JSON: examples/mcp/

Agent frameworks (LangGraph · LlamaIndex · OpenAI Agents)

Connect frameworks to this MCP — do not embed them in the gateway.

Guide	Focus
docs/frameworks/	Index + checklist
docs/frameworks/langgraph.md	LangGraph / LangChain
docs/frameworks/llamaindex.md	LlamaIndex
docs/frameworks/openai-agents.md	OpenAI Agents SDK

Install (npm / npx) — not from this repo

# Hosted HTTP + admin UI (default)
npx -y @markusvankempen/slack-wxo-mcp-gateway

# IDE / stdio MCP
npx -y @markusvankempen/slack-wxo-mcp-gateway --stdio

Requires Node 18+ and Python 3.10+. Env template: .env.example. Guides: local-ngrok · code-engine.

---

Mental model

Multi-channel routing:

flowchart LR
  S1["#support"] --> G["Gateway bindings"]
  S2["#orders"] --> G
  S3["#ops"] --> G
  G --> WA["WxO agent A"]
  G --> WB["WxO agent B"]
  G --> WC["WxO agent C"]

Message path (reply_mode: gateway_thread):

sequenceDiagram
  participant U as Slack user
  participant Ch as Channel / thread
  participant GW as Gateway poller
  participant Wx as WxO Runs API
  participant Bot as Slack bot reply

  U->>Ch: Human message
  GW->>Ch: Read new messages / replies
  GW->>Wx: Start bound agent run
  Wx-->>GW: Agent answer text
  GW->>Bot: chat.postMessage in thread
  Bot-->>Ch: Clean reply (no done noise)

Same host also serves MCP at /mcp and the admin UI at /.

---

Config (config.yaml)

Field	Meaning
slack_channel_id	e.g. C0BHWEZ7NLC
wxo.agent_id	Target Orchestrate agent
mode	poll | events | both
reply_mode	gateway_thread = gateway posts Slack thread after Runs API; agent_tools = only start agent
poll_sec / lookback_sec	Poller timing

Secrets: use ${ENV_VAR} (loaded from .env).

---

Endpoints

Path	Role
/	Admin UI
/mcp	MCP streamable HTTP
/slack/events	Slack Event Subscriptions
/health	Liveness
/api/logs	Log ring buffer
/api/tools	MCP tool catalog
/api/diagnostics	Slack + WxO checks
/api/poll	One poll cycle
/api/config	Masked JSON / raw YAML

Admin dashboard auth

GATEWAY_ADMIN_USER=admin
GATEWAY_ADMIN_PASSWORD=choose-a-strong-password

Protects / and /api/*. Public: /health, /mcp, /slack/events.

IBM Code Engine

./deploy_code_engine.sh
./test_code_engine.sh

Register the toolkit:

orchestrate toolkits add -k mcp -n slack_wxo_gateway \
  --url "https://YOUR-HOST/mcp" \
  --transport streamable_http \
  --tools "*"

MCP tools (14)

Config: list_bindings, upsert_binding
Slack: list_slack_channels, list_recent_messages, list_thread_replies, get_message_context, post_thread_reply, set_typing_indicator
WxO: list_wxo_agents, invoke_wxo_agent
Ops: poll_once, get_gateway_status, get_recent_logs, run_diagnostics_tool

Bot scopes: channels:read, groups:read, reactions:write (reinstall Slack app after adding).

Agents:

Agent	Role
agent.yaml → slack_gateway_test_agent	Full-toolkit smoke
agents/slack_gateway_ops_agent.yaml	Day-2 ops / routing
agents/slack_gateway_answer_agent.yaml	Channel answers (gateway_thread)

Setup (Slack + WxO): SETUP.md — also live in admin UI → Setup
Use cases + test plan: USE_CASES.md
Publish (npm / GitHub): PUBLISH.md

---

Reply modes

gateway_thread (default) — poller/Events → Runs API → gateway chat.postMessage in thread. Use the answer-only agent (no done).

agent_tools — gateway only starts the agent; agent uses its own Slack tools.

---

Cursor / VS Code / Bob / Antigravity / Claude

See docs/ide/ for each client. Quick remote bridge:

{
  "mcpServers": {
    "slack-wxo-gateway": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://YOUR-HOST/mcp"]
    }
  }
}

Package identity:

• npm: @markusvankempen/slack-wxo-mcp-gateway
• MCP: io.github.markusvankempen/slack-wxo-mcp-gateway
• Site: https://markusvankempen.github.io/

---

License

Apache-2.0 — © Markus van Kempen
https://markusvankempen.github.io/ · https://github.com/markusvankempen